SpanBERT: Improving Pre-training by Representing

and Predicting Spans

Mandar Joshi∗† Danqi Chen∗‡§ Yinhan Liu§

Daniel S. Weld† Luke Zettlemoyer†§ Omer Levy§

†
Allen School of Computer Science & Engineering, University of Washington, Seattle, WA
{mandar90,weld,lsz}@cs.washington.edu
‡
Computer Science Department, Princeton University, Princeton, NJ
danqic@cs.princeton.edu

Allen Institute of Artificial Intelligence, Seattle
{danw}@allenai.org

Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00300/1923170/tacl_a_00300.pdf by guest on 16 December 2022

§
Facebook AI Research, Seattle
{danqi,yinhanliu,lsz,omerlevy}@fb.com

Abstract termining that the ‘‘Denver Broncos’’ is a type of

We present SpanBERT, a pre-training method
that is designed to better represent and predict
spans of text. Our approach extends BERT
by (1) masking contiguous random spans,
rather than random tokens, and (2) training
the span boundary representations to predict
the entire content of the masked span, without
relying on the individual token representations
within it. SpanBERT consistently outperforms
BERT and our better-tuned baselines, with
substantial gains on span selection tasks such
as question answering and coreference reso-
lution. In particular, with the same training data
and model size as BERTlarge , our single model
obtains 94.6% and 88.7% F1 on SQuAD 1.1
and 2.0 respectively. We also achieve a new
state of the art on the OntoNotes coreference
resolution task (79.6% F1), strong perfor-
mance on the TACRED relation extraction
benchmark, and even gains on GLUE.1
1 Introduction
Pre-training methods like BERT (Devlin et al.,
2019) have shown strong performance gains using
self-supervised training that masks individual words
or subword units. However, many NLP tasks in-
volve reasoning about relationships between two
or more spans of text. For example, in extractive
question answering (Rajpurkar et al., 2016), de-
∗
Equal contribution.
1
Our code and pre-trained models are available at https://
github.com/facebookresearch/SpanBERT.
64
‘‘NFL team’’ is critical for answering the ques-
tion ‘‘Which NFL team won Super Bowl 50?’’
Such spans provide a more challenging target
for self supervision tasks, for example, predicting
‘‘Denver Broncos’’ is much harder than predicting
only ‘‘Denver’’ when you know the next word is
‘‘Broncos’’. In this paper, we introduce a span-
level pretraining approach that consistently out-
performs BERT, with the largest gains on span
selection tasks such as question answering and
coreference resolution.
We present SpanBERT, a pre-training method
that is designed to better represent and predict
spans of text. Our method differs from BERT in
both the masking scheme and the training objec-
tives. First, we mask random contiguous spans,
rather than random individual tokens. Second, we
introduce a novel span-boundary objective (SBO)
so the model learns to predict the entire masked
span from the observed tokens at its boundary.
Span-based masking forces the model to predict
entire spans solely using the context in which
they appear. Furthermore, the SBO encourages
the model to store this span-level information at
the boundary tokens, which can be easily accessed
during the fine-tuning stage. Figure 1 illustrates
our approach.
To implement SpanBERT, we build on a well-
tuned replica of BERT, which itself substantially
outperforms the original BERT. While building on
our baseline, we find that pre-training on single
segments, instead of two half-length segments
with the next sentence prediction (NSP) objective,

Transactions of the Association for Computational Linguistics, vol. 8, pp. 64–77, 2020. https://doi.org/10.1162/tacl a 00300
Action Editor: Radu Florian. Submission batch: 9/2019; Revision batch: 10/2019; Published 3/2020.
c 2020 Association for Computational Linguistics. Distributed under a CC-BY 4.0 license.

Figure 1: An illustration of SpanBERT training. The span an American football game is masked. The SBO uses
the output representations of the boundary tokens, x4 and x9 (in blue), to predict each token in the masked span.
The equation shows the MLM and SBO loss terms for predicting the token, football (in pink), which as marked
by the position embedding p3 , is the third token from x4 .
considerably improves performance on most
downstream tasks. Therefore, we add our modifi-
cations on top of the tuned single-sequence BERT
baseline.
Together, our pre-training process yields mod-
els that outperform all BERT baselines on a
wide variety of tasks, and reach substantially
better performance on span selection tasks in par-
ticular. Specifically, our method reaches 94.6%
and 88.7% F1 on SQuAD 1.1 and 2.0 (Rajpurkar
et al., 2016, 2018), respectively—reducing error
by as much as 27% compared with our tuned
BERT replica. We also observe similar gains
on five additional extractive question answering
benchmarks (NewsQA, TriviaQA, SearchQA,
HotpotQA, and Natural Questions).2
SpanBERT also arrives at a new state of the art
on the challenging CoNLL-2012 (‘‘OntoNotes’’)
shared task for document-level coreference resolu-
tion, where we reach 79.6% F1, exceeding the pre-
vious top model by 6.6% absolute. Finally,
we demonstrate that SpanBERT also helps on
tasks that do not explicitly involve span selec-
tion, and show that our approach even im-
proves performance on TACRED (Zhang et al.,
2017) and GLUE (Wang et al., 2019).
Whereas others show the benefits of adding
more data (Yang et al., 2019) and increasing
model size (Lample and Conneau, 2019), this
work demonstrates the importance of designing
2
We use the modified MRQA version of these datasets.
See more details in Section 4.1.
65
Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00300/1923170/tacl_a_00300.pdf by guest on 16 December 2022
good pre-training tasks and objectives, which can
also have a remarkable impact.
2 Background: BERT
BERT (Devlin et al., 2019) is a self-supervised
approach for pre-training a deep transformer en-
coder (Vaswani et al., 2017), before fine-tuning
it for a particular downstream task. BERT opti-
mizes two training objectives—masked language
model (MLM) and next sentence prediction (NSP)—
which only require a large collection of unlabeled
text.
Notation Given a sequence of word or sub-
word tokens X = (x1 , x2 , . . . , xn ), BERT trains
an encoder that produces a contextualized vector
representation for each token:
enc(x1 , x2 , . . . , xn ) = x1 , x2 , . . . , xn .
Masked Language Model Also known as a
cloze test, MLM is the task of predicting missing
tokens in a sequence from their placeholders.
Specifically, a subset of tokens Y ⊆ X is sampled
and substituted with a different set of tokens. In
BERT’s implementation, Y accounts for 15% of
the tokens in X ; of those, 80% are replaced with
[MASK], 10% are replaced with a random token
(according to the unigram distribution), and 10%
are kept unchanged. The task is to predict the
original tokens in Y from the modified input.
BERT selects each token in Y independently
by randomly selecting a subset. In SpanBERT, we
define Y by randomly selecting contiguous spans
(Section 3.1).
Next Sentence Prediction The NSP task takes
two sequences (XA , XB ) as input, and predicts
whether XB is the direct continuation of XA . This
is implemented in BERT by first reading XA from
the corpus, and then (1) either reading XB from the
point where XA ended, or (2) randomly sampling
XB from a different point in the corpus. The two
sequences are separated by a special [SEP] token.
Additionally, a special [CLS] token is added to
XA , XB to form the input, where the target of
[CLS] is whether XB indeed follows XA in the
corpus.
In summary, BERT optimizes the MLM and the
NSP objectives by masking word pieces uniformly
at random in data generated by the bi-sequence
sampling procedure. In the next section, we will
present our modifications to the data pipeline,
masking, and pre-training objectives.
3 Model
We present SpanBERT, a self-supervised pre-
training method designed to better represent and
predict spans of text. Our approach is inspired
by BERT (Devlin et al., 2019), but deviates from
its bi-text classification framework in three ways.
First, we use a different random process to mask
spans of tokens, rather than individual ones. We
also introduce a novel auxiliary objective—the
SBO—which tries to predict the entire masked
span using only the representations of the tokens at
the span’s boundary. Finally, SpanBERT samples
a single contiguous segment of text for each train-
ing example (instead of two), and thus does not
use BERT’s next sentence prediction objective,
which we omit.
3.1 Span Masking
Given a sequence of tokens X = (x1 , x2 , . . . , xn ),
we select a subset of tokens Y ⊆ X by iteratively
sampling spans of text until the masking budget
(e.g., 15% of X ) has been spent. At each iteration,
we first sample a span length (number of words)
from a geometric distribution  ∼ Geo(p), which
is skewed towards shorter spans. We then ran-
domly (uniformly) select the starting point for the
span to be masked. We always sample a sequence
of complete words (instead of subword tokens)
and the starting point must be the beginning of
one word. Following preliminary trials,3 we set
3
We experimented with p = {0.1, 0.2, 0.4} and found 0.2
to perform the best.
66
Figure 2: We sample random span lengths from a
Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00300/1923170/tacl_a_00300.pdf by guest on 16 December 2022
geometric distribution  ∼ Geo(p = 0.2) clipped at
max = 10.
p = 0.2, and also clip  at max = 10. This yields
a mean span length of mean () = 3.8. Figure 2
shows the distribution of span mask lengths.
As in BERT, we also mask 15% of the tokens
in total: replacing 80% of the masked tokens with
[MASK], 10% with random tokens, and 10% with
the original tokens. However, we perform this
replacement at the span level and not for each
token individually; that is, all the tokens in a span
are replaced with [MASK] or sampled tokens.
3.2 Span Boundary Objective
Span selection models (Lee et al., 2016, 2017;
He et al., 2018) typically create a fixed-length
representation of a span using its boundary tokens
(start and end). To support such models, we would
ideally like the representations for the end of the
span to summarize as much of the internal span
content as possible. We do so by introducing a
span boundary objective that involves predicting
each token of a masked span using only the
representations of the observed tokens at the
boundaries (Figure 1).
Formally, we denote the output of the trans-
former encoder for each token in the sequence
by x1 , . . . , xn . Given a masked span of tokens
(xs , . . . , xe ) ∈ Y , where (s, e) indicates its start
and end positions, we represent each token xi
in the span using the output encodings of the
external boundary tokens xs−1 and xe+1 , as well
as the position embedding of the target token
pi−s+1 :
yi = f (xs−1, xe+1 , pi−s+1 )
where position embeddings p1 , p2 , . . . mark rela-
tive positions of the masked tokens with respect
to the left boundary token xs−1 . We implement
the representation function f (·) as a 2-layer
feed-forward network with GeLU activations
(Hendrycks and Gimpel, 2016) and layer normal-
ization (Ba et al., 2016):
h0 = [xs−1 ; xe+1 ; pi−s+1 ]
h1 = LayerNorm (GeLU(W1 h0 ))
yi = LayerNorm (GeLU(W2 h1 ))
We then use the vector representation yi to predict
the token xi and compute the cross-entropy loss
exactly like the MLM objective.
SpanBERT sums the loss from both the span
boundary and the regular masked language model
objectives for each token xi in the masked span
(xs , . . . , xe ), while reusing the input embedding
(Press and Wolf, 2017) for the target tokens in
both MLM and SBO:
L(xi ) = LMLM (xi ) + LSBO (xi )
= − log P (xi | xi ) − log P (xi | yi )
3.3 Single-Sequence Training
As described in Section 2, BERT’s examples con-
tain two sequences of text (XA , XB ), and an
objective that trains the model to predict whether
they are connected (NSP). We find that this set-
ting is almost always worse than simply using a
single sequence without the NSP objective (see
Section 5 for further details). We conjecture that
single-sequence training is superior to bi-sequence
training with NSP because (a) the model benefits
from longer full-length contexts, or (b) condi-
tioning on, often unrelated, context from an-
other document adds noise to the masked language
model. Therefore, in our approach, we remove
both the NSP objective and the two-segment sam-
pling procedure, and simply sample a single con-
tiguous segment of up to n = 512 tokens, rather
than two half-segments that sum up to n tokens
together.
In summary, SpanBERT pre-trains span repre-
sentations by: (1) masking spans of full words
using a geometric distribution based masking
scheme (Section 3.1), (2) optimizing an auxiliary
span-boundary objective (Section 3.2) in addition
67
to MLM using a single-sequence data pipeline
(Section 3.3). A procedural description can be
found in Appendix A.
4 Experimental Setup
4.1 Tasks
We evaluate on a comprehensive suite of tasks,
including seven question answering tasks, coref-
erence resolution, nine tasks in the GLUE bench-
mark (Wang et al., 2019), and relation extraction.
We expect that the span selection tasks, question
answering and coreference resolution, will partic-
ularly benefit from our span-based pre-training.
Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00300/1923170/tacl_a_00300.pdf by guest on 16 December 2022
Extractive Question Answering Given a short
passage of text and a question as input, the task
of extractive question answering is to select a
contiguous span of text in the passage as the
answer.
We first evaluate on SQuAD 1.1 and 2.0
(Rajpurkar et al., 2016, 2018), which have
served as major question answering benchmarks,
particularly for pre-trained models (Peters et al.,
2018; Devlin et al., 2019; Yang et al., 2019).
We also evaluate on five more datasets from
the MRQA shared task (Fisch et al., 2019)4 :
NewsQA (Trischler et al., 2017), SearchQA
(Dunn et al., 2017), TriviaQA (Joshi et al.,
2017), HotpotQA (Yang et al., 2018), and Natural
Questions (Kwiatkowski et al., 2019). Because
the MRQA shared task does not have a public
test set, we split the development set in half
to make new development and test sets. The
datasets vary in both domain and collection meth-
odology, making this collection a good test bed
for evaluating whether our pre-trained models can
generalize well across different data distributions.
Following BERT (Devlin et al., 2019), we use
the same QA model architecture for all the data-
sets. We first convert the passage P = (p1 , p2 ,
. . . , pl ) and question Q = (q1 , q2 , . . . , ql ) into
a single sequence X = [CLS]p1 p2 . . . pl [SEP]
q1 q2 . . . ql [SEP], pass it to the pre-trained trans-
former encoder, and train two linear classifiers
independently on top of it for predicting the answer
span boundary (start and end). For the unanswer-
able questions in SQuAD 2.0, we simply set the
4
https://github.com/mrqa/MRQA-Shared-
Task-2019. MRQA changed the original datasets to unify
them into the same format, e.g., all the contexts are truncated
to a maximum of 800 tokens and only answerable questions
are kept.
answer span to be the special token [CLS] for
both training and testing.
Coreference Resolution Coreference resolu-
tion is the task of clustering mentions in text which
refer to the same real-world entities. We evaluate
on the CoNLL-2012 shared task (Pradhan et al.,
2012) for document-level coreference resolu-
tion. We use the independent version of the Joshi
et al. (2019b) implementation of the higher-order
coreference model (Lee et al., 2018). The docu-
ment is divided into non-overlapping segments of
a pre-defined length.5 Each segment is encoded
independently by the pre-trained transformer
encoder, which replaces the original LSTM-based
encoder. For each mention span x, the model
learns a distribution P (·) over possible antecedent
spans Y :
es(x,y)
P (y ) =  s(x,y  )
y  ∈Y e
The span pair scoring function s(x, y ) is a feed-
forward neural network over fixed-length span
representations and hand-engineered features over
x and y :
s(x, y ) = sm (x) + sm (y ) + sc (x, y )
sm (x) = FFNN m (gx )
sc (x, y ) = FFNN c (gx , gy , φ(x, y ))
Here gx and gy denote the span representations,
which are a concatenation of the two transformer
output states of the span endpoints and an attention
vector computed over the output representations
of the token in the span. FFNNm and FFNNc
represent two feedforward neural networks with
one hidden layer, and φ(x, y ) represents the hand-
engineered features (e.g., speaker and genre infor-
mation). A more detailed description of the model
can be found in Joshi et al. (2019b).
Relation Extraction TACRED (Zhang et al.,
2017) is a challenging relation extraction dataset.
Given one sentence and two spans within it—
subject and object—the task is to predict the
relation between the spans from 42 pre-defined
relation types, including no relation. We follow
the entity masking schema from Zhang et al.
(2017) and replace the subject and object entities
by their NER tags such as ‘‘[CLS] [SUBJ-PER]
The length was chosen from {128, 256, 384, 512}. See
5
more details in Appendix B.
68
was born in [OBJ-LOC] , Michigan, . . . ’’, and
finally add a linear classifier on top of the [CLS]
token to predict the relation type.
GLUE The General Language Understanding
Evaluation (GLUE) benchmark (Wang et al.,
2019) consists of 9 sentence-level classification
tasks:
• Two sentence-level classification tasks in-
cluding CoLA (Warstadt et al., 2018)
for evaluating linguistic acceptability and
SST-2 (Socher et al., 2013) for sentiment
classification.
Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00300/1923170/tacl_a_00300.pdf by guest on 16 December 2022
• Three sentence-pair similarity tasks includ-
ing MRPC (Dolan and Brockett, 2005), a
binary paraphrasing task sentence pairs from
news sources, STS-B (Cer et al., 2017), a
graded similarity task for news headlines,
and QQP,6 a binary paraphrasing tasking be-
tween Quora question pairs.
• Four natural language inference tasks in-
cluding MNLI (Williams et al., 2018), QNLI
(Rajpurkar et al., 2016), RTE (Dagan et al.,
2005; Bar-Haim et al., 2006; Giampiccolo
et al., 2007), and WNLI (Levesque et al.,
2011).
Unlike question answering, coreference resolu-
tion, and relation extraction, these sentence-level
tasks do not require explicit modeling of span-
level semantics. However, they might still benefit
from implicit span-based reasoning (e.g., the
Prime Minister is the head of the government).
Following previous work (Devlin et al., 2019;
Radford et al., 2018),7 we exclude WNLI from
the results to enable a fair comparison. Although
recent work Liu et al. (2019a) has applied several
task-specific strategies to increase performance
on the individual GLUE tasks, we follow BERT’s
single-task setting and only add a linear classi-
fier on top of the [CLS] token for these classifi-
cation tasks.
4.2 Implementation
We reimplemented BERT’s model and pre-
training method in fairseq (Ott et al., 2019).
6
https://data.quora.com/First-Quora-
Dataset-Release-Question-Pairs.
7
Previous work has excluded WNLI on account of con-
struction issues outlined on the GLUE website – https://
gluebenchmark.com/faq.
We used the model configuration of BERTlarge SQuAD 1.1 SQuAD 2.0
as in Devlin et al. (2019) and also pre-trained all EM F1 EM F1
our models on the same corpus: BooksCorpus and
Human Perf. 82.3 91.2 86.8 89.4
English Wikipedia using cased Wordpiece tokens.
Google BERT 84.3 91.3 80.0 83.3
Compared with the original BERT implemen- Our BERT 86.5 92.6 82.8 85.9
tation, the main differences in our implementation Our BERT-1seq 87.5 93.3 83.8 86.6
include: (a) We use different masks at each epoch SpanBERT 88.8 94.6 85.7 88.7
while BERT samples 10 different masks for each
sequence during data processing. (b) We remove Table 1: Test results on SQuAD 1.1 and SQuAD
all the short-sequence strategies used before (they 2.0.
sampled shorter sequences with a small proba-
bility 0.1; they also first pre-trained with smaller Our BERT-1seq Our reimplementation of BERT
sequence length of 128 for 90% of the steps). trained on single full-length sequences without

Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00300/1923170/tacl_a_00300.pdf by guest on 16 December 2022

Instead, we always take sequences of up to 512 NSP (Section 3.3).
tokens until it reaches a document boundary. We
refer readers to Liu et al. (2019b) for further dis- 5 Results
cussion on these modifications and their effects.
We compare SpanBERT to the baselines per task,
As in BERT, the learning rate is warmed up
and draw conclusions based on the overall trends.
over the first 10,000 steps to a peak value of 1e-4,
and then linearly decayed. We retain β hyper- 5.1 Per-Task Results
parameters (β1 = 0.9, β2 = 0.999) and a de-
Extractive Question Answering Table 1 shows
coupled weight decay (Loshchilov and Hutter,
the performance on both SQuAD 1.1 and 2.0.
2019) of 0.1. We also keep a dropout of 0.1 on
SpanBERT exceeds our BERT baseline by 2.0%
all layers and attention weights, and a GeLU ac-
and 2.8% F1, respectively (3.3% and 5.4% over
tivation function (Hendrycks and Gimpel, 2016).
Google BERT). In SQuAD 1.1, this result ac-
We deviate from the optimization by running
counts for over 27% error reduction, reaching
for 2.4M steps and using an epsilon of 1e-8 for
3.4% F1 above human performance.
AdamW (Kingma and Ba, 2015), which con-
Table 2 demonstrates that this trend goes be-
verges to a better set of model parameters. Our im-
yond SQuAD, and is consistent in every MRQA
plementation uses a batch size of 256 sequences
dataset. On average, we see a 2.9% F1 improve-
with a maximum of 512 tokens.8 For the SBO,
ment from our reimplementation of BERT. Al-
we use 200 dimension position embeddings p1 ,
though some gains are coming from single-sequence
p2 , . . . to mark positions relative to the left
training (+1.1%), most of the improvement stems
boundary token. The pre-training was done on 32
from span masking and the span boundary objec-
Volta V100 GPUs and took 15 days to complete.
tive (+1.8%), with particularly large gains on
Fine-tuning is implemented based on Hugging-
TriviaQA (+3.2%) and HotpotQA (+2.7%).
Face’s codebase (Wolf et al., 2019) and more
details are given in Appendix B. Coreference Resolution Table 3 shows the
performance on the OntoNotes coreference res-
4.3 Baselines olution benchmark. Our BERT reimplementation
improves the Google BERT model by 1.2% on
We compare SpanBERT to three baselines:
the average F1 metric and single-sequence train-
Google BERT The pre-trained models released ing brings another 0.5% gain. Finally, SpanBERT
by Devlin et al. (2019).9 improves considerably on top of that, achieving a
new state of the art of 79.6% F1 (previous best
Our BERT Our reimplementation of BERT result is 73.0%).
with improved data preprocessing and optimiza- Relation Extraction Table 4 shows the perfor-
tion (Section 4.2). mance on TACRED. SpanBERT exceeds our
8
On the average, this is approximately 390 sequences,
reimplementation of BERT by 3.3% F1 and
because some documents have fewer than 512 tokens. achieves close to the current state of the art (Soares
9
https://github.com/google-research/bert. et al., 2019)—Our model performs better than

69
 NewsQA TriviaQA SearchQA HotpotQA Natural Avg.
Questions
Google BERT 68.8 77.5 81.7 78.3 79.9 77.3
Our BERT 71.0 79.0 81.8 80.5 80.5 78.6
Our BERT-1seq 71.9 80.4 84.0 80.3 81.8 79.7
SpanBERT 73.6 83.6 84.8 83.0 82.5 81.5

Table 2: Performance (F1) on the five MRQA extractive question answering tasks.

MUC B3 CEAFφ4
P R F1 P R F1 P R F1 Avg. F1
Prev. SotA:
(Lee et al., 2018) 81.4 79.5 80.4 72.2 69.5 70.8 68.2 67.1 67.6 73.0

Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00300/1923170/tacl_a_00300.pdf by guest on 16 December 2022

Google BERT 84.9 82.5 83.7 76.7 74.2 75.4 74.6 70.1 72.3 77.1
Our BERT 85.1 83.5 84.3 77.3 75.5 76.4 75.0 71.9 73.9 78.3
Our BERT-1seq 85.5 84.1 84.8 77.8 76.7 77.2 75.3 73.5 74.4 78.8
SpanBERT 85.8 84.8 85.3 78.3 77.9 78.1 76.4 74.2 75.3 79.6

Table 3: Performance on the OntoNotes coreference resolution benchmark. The main evaluation is the
average F1 of three metrics: MUC, B3 , and CEAFφ4 on the test set.

p R F1 training without the NSP objective substantially

improves CoLA, and yields smaller (but consid-
BERTEM (Soares et al., 2019) − − 70.1
erable) improvements on MRPC and MNLI. The
BERTEM +MTB∗ − − 71.5
main gains from SpanBERT are in the SQuAD-
Google BERT 69.1 63.9 66.4 based QNLI dataset (+1.3%) and in RTE (+6.9%),
Our BERT 67.8 67.2 67.5 the latter accounting for most of the rise in
Our BERT-1seq 72.4 67.9 70.1 SpanBERT’s GLUE average.
SpanBERT 70.8 70.9 70.8
5.2 Overall Trends
Table 4: Test performance on the TACRED
relation extraction benchmark. BERTlarge and We compared our approach to three BERT base-
BERTEM +MTB from Soares et al. (2019) are the lines on 17 benchmarks, and found that SpanBERT
current state-of-the-art. ∗ : BERTEM +MTB incor- outperforms BERT on almost every task. In 14
porated an intermediate ‘‘matching the blanks’’ tasks, SpanBERT performed better than all base-
pre-training on the entity-linked text based on lines. In two tasks (MRPC and QQP), it performed
English Wikipedia, which is not a direct compar- on-par in terms of accuracy with single-sequence
ison to ours trained only from raw text. trained BERT, but still outperformed the other
baselines. In one task (SST-2), Google’s BERT
baseline performed better than SpanBERT by
their BERTEM but is 0.7 point behind BERTEM + 0.4% accuracy.
MTB, which used entity-linked text for additional
When considering the magnitude of the gains,
pre-training. Most of this gain (+2.6%) stems from
it appears that SpanBERT is especially better at
single-sequence training although the contribution
extractive question answering. In SQuAD 1.1,
of span masking and the span boundary objective
for example, we observe a solid gain of 2.0% F1
is still a considerable 0.7%, resulting largely from
even though the baseline is already well above
higher recall.
human performance. On MRQA, SpanBERT im-
GLUE Table 5 shows the performance on proves between 2.0% (Natural Questions) and
GLUE. 4.6% (TriviaQA) F1 on top of our BERT baseline.
For most tasks, the different models appear Finally, we observe that single-sequence train-
to perform similarly. Moving to single-sequence ing works considerably better than bi-sequence

70
 CoLA SST-2 MRPC STS-B
Google BERT 59.3 95.2 88.5/84.3 86.4/88.0
Our BERT 58.6 93.9 90.1/86.6 88.4/89.1
Our BERT-1seq 63.5 94.8 91.2/87.8 89.0/88.4
SpanBERT 64.3 94.8 90.9/87.9 89.9/89.1
Table 5: Test set performance on GLUE tasks. MRPC: F1/accuracy, STS-B: Pearson/Spearmanr
correlation, QQP: F1/accuracy, MNLI: matched/mistached accuracies, and accuracy for all the other
tasks. WNLI (not shown) is always set to majority class (65.1% accuracy) and included in the average.
training with NSP with BERT’s choice of se-
quence lengths for a wide variety of tasks. This
is surprising because BERT’s ablations showed
gains from the NSP objective (Devlin et al., 2019).
However, the ablation studies still involved bi-
sequence data processing (i.e., the pre-training
stage only controlled for the NSP objective while
still sampling two half-length sequences). We hy-
pothesize that bi-sequence training, as it is im-
plemented in BERT (see Section 2), impedes the
model from learning longer-range features, and
consequently hurts performance on many down-
stream tasks.
6 Ablation Studies
We compare our random span masking scheme
with linguistically-informed masking schemes,
and find that masking random spans is a com-
petitive and often better approach. We then study
the impact of the SBO, and contrast it with BERT’s
NSP objective.10
6.1 Masking Schemes
Previous work (Sun et al., 2019) has shown im-
provements in downstream task performance by
masking linguistically informed spans during pre-
training for Chinese data. We compare our ran-
dom span masking scheme with masking of
linguistically informed spans. Specifically, we
train the following five baseline models differing
only in the way tokens are masked.
Subword Tokens We sample random Word-
piece tokens, as in the original BERT.
Whole Words We sample random words, and
then mask all of the subword tokens in those
words. The total number of masked subtokens is
around 15%.
10
To save time and resources, we use the checkpoints at
1.2M steps for all the ablation experiments.
71
QQP MNLI QNLI RTE (Avg)
71.2/89.0 86.1/85.7 93.0 71.1 80.4
71.8/89.3 87.2/86.6 93.0 74.7 81.1
72.1/89.5 88.0/87.4 93.0 72.1 81.7
71.9/89.5 88.1/87.7 94.3 79.0 82.8
Named Entities At 50% of the time, we sample
from named entities in the text, and sample random
whole words for the other 50%. The total number
Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00300/1923170/tacl_a_00300.pdf by guest on 16 December 2022
of masked subtokens is 15%. Specifically, we run
spaCy’s named entity recognizer (Honnibal and
Montani, 2017)11 on the corpus and select all the
non-numerical named entity mentions as candidates.
Noun Phrases Similar to Named Entities, we
sample from noun phrases at 50% of the time. The
noun phrases are extracted by running spaCy’s
constituency parser.
Geometric Spans We sample random spans
from a geometric distribution, as in our SpanBERT
(see Section 3.1).
Table 6 shows how different pre-training
masking schemes affect performance on the devel-
opment set of a selection of tasks. All the mod-
els are evaluated on the development sets and are
based on the default BERT setup of bi-sequence
training with NSP; the results are not directly com-
parable to the main evaluation. With the exception
of coreference resolution, masking random spans
is preferable to other strategies. Although linguis-
tic masking schemes (named entities and noun
phrases) are often competitive with random spans,
their performance is not consistent; for instance,
masking noun phrases achieves parity with ran-
dom spans on NewsQA, but underperforms on
TriviaQA (−1.1% F1).
On coreference resolution, we see that masking
random subword tokens is preferable to any form
of span masking. Nevertheless, we shall see in
the following experiment that combining random
span masking with the span boundary objective
can improve upon this result considerably.
6.2 Auxiliary Objectives
In Section 5, we saw that bi-sequence training
with the NSP objective can hurt performance on
11
https://spacy.io/.
 SQuAD 2.0 NewsQA TriviaQA Coreference MNLI-m QNLI GLUE (Avg)
Subword Tokens 83.8 72.0 76.3 77.7 86.7 92.5 83.2
Whole Words 84.3 72.8 77.1 76.6 86.3 92.8 82.9
Named Entities 84.8 72.7 78.7 75.6 86.0 93.1 83.2
Noun Phrases 85.0 73.0 77.7 76.7 86.5 93.2 83.5
Geometric Spans 85.4 73.0 78.8 76.4 87.0 93.3 83.4

Table 6: The effect of replacing BERT’s original masking scheme (Subword Tokens) with different
masking schemes. Results are F1 scores for QA tasks and accuracy for MNLI and QNLI on the
development sets. All the models are based on bi-sequence training with NSP.

SQuAD 2.0 NewsQA TriviaQA Coref MNLI-m QNLI GLUE (Avg)

Span Masking (2seq) + NSP 85.4 73.0 78.8 76.4 87.0 93.3 83.4

Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00300/1923170/tacl_a_00300.pdf by guest on 16 December 2022

Span Masking (1seq) 86.7 73.4 80.0 76.3 87.3 93.8 83.8
Span Masking (1seq) + SBO 86.8 74.1 80.3 79.0 87.6 93.9 84.0

Table 7: The effects of different auxiliary objectives, given MLM over random spans as the primary
objective.

downstream tasks, when compared with single-
sequence training. We test whether this holds true
for models pre-trained with span masking, and also
evaluate the effect of replacing the NSP objective
with the SBO.
Table 7 confirms that single-sequence training
typically improves performance. Adding SBO fur-
ther improves performance, with a substantial gain
on coreference resolution (+2.7% F1) over span
masking alone. Unlike the NSP objective, SBO
does not appear to have any adverse effects.
7 Related Work
Pre-trained contextualized word representations that
can be trained from unlabeled text (Dai and Le,
2015; Melamud et al., 2016; Peters et al., 2018)
have had immense impact on NLP lately, partic-
ularly as methods for initializing a large model
before fine-tuning it for a specific task (Howard
and Ruder, 2018; Radford et al., 2018; Devlin
et al., 2019). Beyond differences in model hyper-
parameters and corpora, these methods mainly
differ in their pre-training tasks and loss functions,
with a considerable amount of contemporary liter-
ature proposing augmentations of BERT’s MLM
objective.
While previous and concurrent work has looked
at masking (Sun et al., 2019) or dropping (Song
et al., 2019; Chan et al., 2019) multiple words
from the input—particularly as pretraining for lan-
guage generation tasks—SpanBERT pretrains
span representations (Lee et al., 2016), which are
widely used for question answering, coreference
resolution, and a variety of other tasks. ERNIE
(Sun et al., 2019) shows improvements on Chinese
NLP tasks using phrase and named entity mask-
ing. MASS (Song et al., 2019) focuses on language
generation tasks, and adopts the encoder-decoder
framework to reconstruct a sentence fragment
given the remaining part of the sentence. We
attempt to more explicitly model spans using the
SBO objective, and show that (geometrically dis-
tributed) random span masking works as well,
and sometimes better than, masking linguistically-
coherent spans. We evaluate on English bench-
marks for question answering, relation extraction,
and coreference resolution in addition to GLUE.
A different ERNIE (Zhang et al., 2019) fo-
cuses on integrating structured knowledge bases
with contextualized representations with an eye on
knowledge-driven tasks like entity typing and re-
lation classification. UNILM (Dong et al., 2019)
uses multiple language modeling objectives—
unidirectional (both left-to-right and right-to-left),
bidirectional, and sequence-to-sequence prediction—
to aid generation tasks like summarization and
question generation. XLM (Lample and Conneau,
2019) explores cross-lingual pre-training for multi-
lingual tasks such as translation and cross-lingual
classification. Kermit (Chan et al., 2019), an in-
sertion based approach, fills in missing tokens

72
(instead of predicting masked ones) during pre-
training; they show improvements on machine
translation and zero-shot question answering.
Concurrent with our work, RoBERTa (Liu et al.,
2019b) presents a replication study of BERT
pre-training that measures the impact of many
key hyperparameters and training data size. Also
concurrent, XLNet (Yang et al., 2019) combines
an autoregressive loss and the Transformer-XL
(Dai et al., 2019) architecture with a more than an
eight-fold increase in data to achieve current state-
of-the-art results on multiple benchmarks. XLNet
also masks spans (of 1–5 tokens) during pre-
training, but predicts them autoregressively. Our
model focuses on incorporating span-based pre-
training, and as a side effect, we present a stronger
BERT baseline while controlling for the corpus,
architecture, and the number of parameters.
Related to our SBO objective, pair2vec (Joshi
et al., 2019a) encodes word-pair relations using
a negative sampling-based multivariate objective
during pre-training. Later, the word-pair repre-
sentations are injected into the attention-layer of
downstream tasks, and thus encode limited down-
stream context. Unlike pair2vec, our SBO objec-
tive yields ‘‘pair’’ (start and end tokens of spans)
representations which more fully encode the con-
text during both pre-training and finetuning, and
are thus more appropriately viewed as span repre-
sentations. Stern et al. (2018) focus on improving
language generation speed using a block-wise par-
allel decoding scheme; they make predictions for
multiple time steps in parallel and then back off
to the longest prefix validated by a scoring model.
Also related are sentence representation methods
(Kiros et al., 2015; Logeswaran and Lee, 2018),
which focus on predicting surrounding contexts
from sentence embeddings.
8 Conclusion
We presented a new method for span-based pre-
training which extends BERT by (1) masking
contiguous random spans, rather than random
tokens, and (2) training the span boundary repre-
sentations to predict the entire content of the
masked span, without relying on the individual
token representations within it. Together, our pre-
training process yields models that outperform all
BERT baselines on a variety of tasks, and reach
73
substantially better performance on span selection
tasks in particular.
Appendices
A Pre-training Procedure
We describe our pre-training procedure as follows:
1. Divide the corpus into single contiguous
blocks of up to 512 tokens.
2. At each step of pre-training:
(a) Sample a batch of blocks uniformly at
Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00300/1923170/tacl_a_00300.pdf by guest on 16 December 2022
random.
(b) Mask 15% of word pieces in each block
in the batch using the span masking
scheme (Section 3.1).
(c) For each masked token xi , opti-
mize L(xi ) = LMLM (xi ) + LSBO (xi )
(Section 3.2).
B Fine-tuning Hyperparameters
We apply the following fine-tuning hyperparam-
eters to all methods, including the baselines.
Extractive Question Answering For all the
question answering tasks, we use max seq
length = 512 and a sliding window of size
128 if the lengths are longer than 512. We choose
learning rates from {5e-6, 1e-5, 2e-5, 3e-5, 5e-5}
and batch sizes from {16, 32} and fine-tune four
epochs for all the datasets.
Coreference Resolution We divide the docu-
ments into multiple chunks of lengths up to max
seq length and encode each chunk indepen-
dently. We choose max seq length from {128,
256, 384, 512}, BERT learning rates from {1e-5,
2e-5}, task-specific learning rates from {1e-4,
2e-4, 3e-4}, and fine-tune 20 epochs for all the
datasets. We use batch size = 1 (one document)
for all the experiments.
TACRED/GLUE We use max seq length =
128 and choose learning rates from {5e-6, 1e-5,
2e-5, 3e-5, 5e-5} and batch sizes from {16, 32}
and fine-tuning 10 epochs for all the datasets.
The only exception is CoLA, where we used four
epochs (following Devlin et al., 2019), because 10
epochs lead to severe overfitting.
Acknowledgments
We would like to thank Pranav Rajpurkar and Robin
Jia for patiently helping us evaluate SpanBERT
on SQuAD. We thank the anonymous reviewers,
the action editor, and our colleagues at Facebook
AI Research and the University of Washington for
their insightful feedback that helped improve the
paper.
References
Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E.
Hinton. 2016. Layer normalization. arXiv pre-
print arXiv:1607.06450.
Roy Bar-Haim, Ido Dagan, Bill Dolan, Lisa Ferro,
Danilo Giampiccolo, Bernardo Magnini, and
Idan Szpektor. 2006. The second PASCAL
recognising textual entailment challenge. In
Proceedings of the Second PASCAL Challenges
Workshop on Recognising Textual Entailment,
pages 6–4.
Daniel Cer, Mona Diab, Eneko Agirre, IÃśigo
Lopez-Gazpio, and Lucia Specia. 2017.
Semeval-2017 task 1: Semantic textual similar-
ity multilingual and crosslingual focused eval-
uation. In International Workshop on Semantic
Evaluation (SemEval), pages 1–14. Vancouver,
Canada.
William Chan, Nikita Kitaev, Kelvin Guu,
Mitchell Stern, and Jakob Uszkoreit. 2019.
KERMIT: Generative insertion-based modeling
for sequences. arXiv preprint arXiv:1906.01604.
Ido Dagan, Oren Glickman, and Bernardo
Magnini. 2005. The PASCAL recognising tex-
tual entailment challenge. In Machine Learning
Challenges Workshop, pages 177–190. Springer.
Andrew M. Dai and Quoc V. Le. 2015. Semi-
supervised sequence learning. In Advances in
Neural Information Processing Systems (NIPS),
pages 3079–3087.
Zihang Dai, Zhilin Yang, Yiming Yang, William
W. Cohen, Jaime Carbonell, Quoc V. Le, and
Ruslan Salakhutdinov. 2019. Transformer-XL:
Attentive language models beyond a fixed-
length context. In Association for Computa-
tional Linguistics (ACL).
74
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and
Kristina Toutanova. 2019. BERT: Pre-training
of deep bidirectional transformers for language
understanding. In North American Association
for Computational Linguistics (NAACL).
William B. Dolan and Chris Brockett. 2005.
Automatically constructing a corpus of sen-
tential paraphrases. In Proceedings of the Inter-
national Workshop on Paraphrasing.
Li Dong, Nan Yang, Wenhui Wang, Furu Wei,
Xiaodong Liu, Yu Wang, Jianfeng Gao, Ming
Zhou, and Hsiao-Wuen Hon. 2019. Unified
Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00300/1923170/tacl_a_00300.pdf by guest on 16 December 2022
language model pre-training for natural lan-
guage understanding and generation. In Ad-
vances in Neural Information Processing Systems
(NIPS).
Matthew Dunn, Levent Sagun, Mike Higgins,
V. Ugur Guney, Volkan Cirik, and Kyunghyun
Cho. 2017. SearchQA: A new Q&A dataset
augmented with context from a search engine.
arXiv preprint arXiv:1704.05179.
Adam Fisch, Alon Talmor, Robin Jia, Minjoon
Seo, Eunsol Choi, and Danqi Chen. 2019.
MRQA 2019 shared task: Evaluating general-
ization in reading comprehension. In Proceed-
ings of 2nd Machine Reading for Reading
Comprehension (MRQA) Workshop at EMNLP.
Danilo Giampiccolo, Bernardo Magnini, Ido
Dagan, and Bill Dolan. 2007. The third PASCAL
recognizing textual entailment challenge. In Pro-
ceedings of the ACL-PASCAL Workshop on Tex-
tual Entailment and Paraphrasing, pages 1–9.
Luheng He, Kenton Lee, Omer Levy, and Luke
Zettlemoyer. 2018. Jointly predicting predicates
and arguments in neural semantic role labeling.
In Association for Computational Linguistics
(ACL), pages 364–369.
Dan Hendrycks and Kevin Gimpel. 2016.
Gaussian error linear units (gelus). arXiv pre-
print arXiv:1606.08415.
Matthew Honnibal and Ines Montani. 2017. spaCy
2: Natural language understanding with Bloom
embeddings, convolutional neural networks and
incremental parsing. To appear.
Jeremy Howard and Sebastian Ruder. 2018. Uni-
versal language model fine-tuning for text clas-
sification. arXiv preprint arXiv:1801.06146.
Mandar Joshi, Eunsol Choi, Omer Levy,
Daniel Weld, and Luke Zettlemoyer. 2019a.
pair2vec: Compositional word-pair embeddings
for cross-sentence inference. In North American
Association for Computational Linguistics
(NAACL), pages 3597–3608.
Mandar Joshi, Eunsol Choi, Daniel Weld, and
Luke Zettlemoyer. 2017. TriviaQA: A large
scale distantly supervised challenge dataset for
reading comprehension. In Association for Com-
putational Linguistics (ACL), pages 1601–1611.
Mandar Joshi, Omer Levy, Daniel S. Weld, Luke
Zettlemoyer, and Omer Levy. 2019b. BERT for
coreference resolution: Baselines and analysis.
In Empirical Methods in Natural Language
Processing (EMNLP).
Diederik Kingma and Jimmy Ba. 2015. Adam:
A method for stochastic optimization. In Inter-
national Conference on Learning Representa-
tions (ICLR).
Ryan Kiros, Yukun Zhu, Ruslan R. Salakhutdinov,
Richard S. Zemel, Antonio Torralba, Raquel
Urtasun, and Sanja Fidler. 2015. Skip-thought
vectors. In Advances in Neural Information
Processing Systems (NIPS).
Tom Kwiatkowski, Jennimaria Palomaki, Olivia
Redfield, Michael Collins, Ankur Parikh, Chris
Alberti, Danielle Epstein, Illia Polosukhin,
Matthew Kelcey, Jacob Devlin, Kenton Lee,
Kristina N. Toutanova, Llion Jones, Ming-
Wei Chang, Andrew Dai, Jakob Uszkoreit,
Quoc Le, and Slav Petrov. 2019. Natural
questions: A benchmark for question answering
research. Transactions of the Association of
Computational Linguistics (TACL).
Guillaume Lample and Alexis Conneau. 2019.
Cross-lingual language model pretraining.
Advances in Neural Information Processing
Systems (NIPS).
Kenton Lee, Luheng He, Mike Lewis, and
Luke Zettlemoyer. 2017. End-to-end neural
coreference resolution. In Empirical Methods
in Natural Language Processing (EMNLP),
pages 188–197.
Kenton Lee, Luheng He, and Luke Zettlemoyer.
2018. Higher-order coreference resolution with
75
coarse-to-fine inference. In North American
Association for Computational Linguistics
(NAACL), pages 687–692.
Kenton Lee, Shimi Salant, Tom Kwiatkowski,
Ankur Parikh, Dipanjan Das, and Jonathan
Berant. 2016. Learning recurrent span repre-
sentations for extractive question answering.
arXiv preprint arXiv:1611.01436.
Hector J. Levesque, Ernest Davis, and Leora
Morgenstern. 2011. The Winograd schema
challenge. In AAAI Spring Symposium: Logical
Formalizations of Commonsense Reasoning,
Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00300/1923170/tacl_a_00300.pdf by guest on 16 December 2022
volume 46, page 47.
Xiaodong Liu, Pengcheng He, Weizhu Chen, and
Jianfeng Gao. 2019a. Multi-task deep neural
networks for natural language understanding. In
Proceedings of the 57th Annual Meeting of the
Association for Computational Linguistics. As-
sociation for Computational Linguistics (ACL).
Yinhan Liu, Myle Ott, Naman Goyal, Jingfei
Du, Mandar Joshi, Danqi Chen, Omer Levy,
Mike Lewis, Luke Zettlemoyer, and Veselin
Stoyanov. 2019b. RoBERTa: A robustly opti-
mized BERT pretraining approach. arxiv pre-
print arXiv:1907.11692.
Lajanugen Logeswaran and Honglak Lee. 2018.
An efficient framework for learning sentence
representations. arxiv preprint arXiv:1803.02893.
Ilya Loshchilov and Frank Hutter. 2019. Decou-
pled weight decay regularization. In Interna-
tional Conference on Learning Representations
(ICLR).
Oren Melamud, Jacob Goldberger, and Ido Dagan.
2016. context2vec: Learning generic context
embedding with bidirectional LSTM. In Com-
putational Natural Language Learning (CoNLL),
pages 51–61.
Myle Ott, Sergey Edunov, Alexei Baevski, Angela
Fan, Sam Gross, Nathan Ng, David Grangier,
and Michael Auli. 2019. fairseq: A fast, exten-
sible toolkit for sequence modeling. In North
American Association for Computational Lin-
guistics (NAACL), pages 48–53.
Matthew Peters, Mark Neumann, Mohit Iyyer,
Matt Gardner, Christopher Clark, Kenton Lee,
 and Luke Zettlemoyer. 2018. Deep contextual-
ized word representations. In North American
Association for Computational Linguistics
(NAACL), pages 2227–2237.
Sameer Pradhan, Alessandro Moschitti, Nianwen
Xue, Olga Uryupina, and Yuchen Zhang. 2012.
CoNLL-2012 shared task: Modeling multi-
lingual unrestricted coreference in ontonotes.
In Joint Conference on EMNLP and CoNLL-
Shared Task, pages 1–40.
Ofir Press and Lior Wolf. 2017. Using the out-
put embedding to improve language models.
In Proceedings of the 15th Conference of the
European Chapter of the Association for Com-
putational Linguistics: Volume 2, Short Papers,
pages 157–163. Association for Computational
Linguistics (ACL).
Alec Radford, Karthik Narasimhan, Time Salimans,
and Ilya Sutskever. 2018. Improving language un-
derstanding with unsupervisedlearning, OpenAI.
Pranav Rajpurkar, Robin Jia, and Percy Liang.
2018. Know what you don’t know: Unanswer-
able questions for SQuAD. In Association for
Computational Linguistics (ACL), pages 784–789.
Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev,
and Percy Liang. 2016. SQuAD: 100,000+
questions for machine comprehension of text.
In Empirical Methods in Natural Language
Processing (EMNLP), pages 2383–2392.
Livio Baldini Soares, Nicholas Arthur FitzGerald,
Jeffrey Ling, and Tom Kwiatkowski. 2019.
Matching the blanks: Distributional similarity
for relation learning. In Association for Compu-
tational Linguistics (ACL), pages 2895–2905.
Richard Socher, Alex Perelygin, Jean Wu, Jason
Chuang, Christopher D. Manning, Andrew
Ng, and Christopher Potts. 2013. Recursive
deep models for semantic compositionality over
a sentiment treebank. In Empirical Methods
in Natural Language Processing (EMNLP),
pages 1631–1642.
Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu,
and Tie-Yan Liu. 2019. MASS: Masked se-
quence to sequence pre-training for language
generation. In International Conference on
Machine Learning (ICML), pages 5926–5936.
76
Mitchell Stern, Noam Shazeer, and Jakob
Uszkoreit. 2018. Blockwise parallel decod-
ing for deep autoregressive models. In Advances
in Neural Information Processing Systems
(NIPS).
Yu Stephanie Sun, Shuohuan Wang, Yukun Li,
Shikun Feng, Xuyi Chen, Han Zhang, Xinlun
Tian, Danxiang Zhu, Hao Tian, and Hua
Wu. 2019. ERNIE: Enhanced representation
through knowledge integration. arXiv preprint
arXiv:1904.09223.
Adam Trischler, Tong Wang, Xingdi Yuan, Justin
Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00300/1923170/tacl_a_00300.pdf by guest on 16 December 2022
Harris, Alessandro Sordoni, Philip Bachman,
and Kaheer Suleman. 2017. NewsQA: A ma-
chine comprehension dataset. In 2nd Work-
shop on Representation Learning for NLP,
pages 191–200.
Ashish Vaswani, Noam Shazeer, Niki Parmar,
Jakob Uszkoreit, Llion Jones, Aidan N. Gomez,
Łukasz Kaiser, and Illia Polosukhin. 2017.
Attention is all you need. In Advances in Neural
Information Processing Systems (NIPS).
Alex Wang, Amapreet Singh, Julian Michael,
Felix Hill, Omer Levy, and Samuel R. Bowman.
2019. GLUE: A multi-task benchmark and anal-
ysis platform for natural language understand-
ing. In International Conference on Learning
Representations (ICLR).
Alex Warstadt, Amanpreet Singh, and Samuel R.
Bowman. 2018. Neural network acceptability
judgments. arXiv preprint arXiv:1805.12471.
Adina Williams, Nikita Nangia, and Samuel
Bowman. 2018. A broad-coverage challenge
corpus for sentence understanding through in-
ference. In North American Association for Com-
putational Linguistics (NAACL),pages1112–1122.
Thomas Wolf, Lysandre Debut, Victor Sanh,
Julien Chaumond, Clement Delangue, Anthony
Moi, Pierric Cistac, Tim Rault, R’emi Louf,
Morgan Funtowicz, and Jamie Brew. 2019.
HuggingFace’s Transformers: State-of-the-art
natural language processing. arXiv preprint
arXiv:1910.03771.
Zhilin Yang, Zihang Dai, Yiming Yang, Jaime
Carbonell, Ruslan Salakhutdinov, and Quoc V.
Le. 2019. XLNet: Generalized autoregressive
 pretraining for language understanding. In
Advances in Neural Information Processing
Systems (NeurIPS).
Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua
Bengio, William Cohen, Ruslan Salakhutdinov,
and Christopher D. Manning. 2018. HotpotQA:
A dataset for diverse, explainable multi-hop
question answering. In Empirical Methods
in Natural Language Processing (EMNLP),
pages 2369–2380.
77
Yuhao Zhang, Victor Zhong, Danqi Chen, Gabor
Angeli, and Christopher D. Manning. 2017.
Position-aware attention and supervised data
improve slot filling. In Empirical Methods
in Natural Language Processing (EMNLP),
pages 35–45.
Zhengyan Zhang, Xu Han, Zhiyuan Liu, Xin
Jiang, Maosong Sun, and Qun Liu. 2019.
ERNIE: Enhanced language representation with
informative entities. In Association for Compu-
tational Linguistics (ACL), pages 1441–1451.
Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00300/1923170/tacl_a_00300.pdf by guest on 16 December 2022